DL-Diff: Advancing Low-Light Video Enhancement through Dark-to-Light Diffusion Models

Key Points

• The study aims to address detail recovery and temporal consistency issues in low-light video enhancement using DL-Diff.
• Developed DL-Diff based on a pre-trained diffusion model for conditional video generation.
• Created cooperative restoration and temporal components to map low-light to normal-light videos and ensure frame consistency.
• Implemented a multi-stage training process to optimize network parameters progressively.
• DL-Diff achieved optimal performance with a FID of 41.29 and LPIPS of 0.17 on the DID dataset.
• Temporal metrics improved significantly with AB (Var) of 25.40 and MABD of 0.08, surpassing comparison LLVE methods.
• Qualitative results showed DL-Diff generates videos with better spatial alignment and temporal coherence than competing methods.